Measuring and Improving the Faithfulness of Model-Generated Reasoning — AI Alignment Forum